Study Evaluates Performance of Machine Learning Algorithm for Automated Detection of Diabetic Retinopathy
Algorithm had high sensitivity and specificity for the identification of diabetic retinopathy and other related eye diseases
Diabetic retinopathy, caused by damage to the blood vessels of the retina, affects the eyes of people with diabetes and can cause blindness if left untreated. An automated solution to diabetic retinopathy detection could lead to faster and cheaper screening, and reach areas where patients currently lack access to such screening programs.
Now, an international team of researchers have developed and trained a machine learning algorithm to identify diabetic retinopathy from retinal photographs. The algorithm, based on a branch of machine learning called deep learning, used a total of 494,661 retinal images from the Singapore National Diabetic Retinopathy Screening Program (SIDRP) as a training set. The study population was highly multiethnic and included Chinese, Indian, Malay, Hispanic, African-American and white patients. Images had labels of referable diabetic retinopathy, possible glaucoma, age-related macular degeneration, or control.
The researchers evaluated the performance of the deep learning-based algorithm with a primary validation dataset of 71,896 images. The sensitivity and specificity of the algorithm were 90.5 percent and 91.6 percent, respectively, for referable diabetic retinopathy. It showed promising results for other eye diseases as well, and the performance was comparable to that of human graders trained by a retinal specialist. Although the trained graders had a higher specificity for referable diabetic retinopathy, the algorithm had higher sensitivity when it came to vision-threatening diabetic retinopathy.
The results were published by JAMA in December.